Stripe Deploys Autonomous Coding Agents at Scale
Stripe's engineering organization has deployed a fleet of autonomous coding agents that now generate 1,300 pull requests per week. The agents handle planning, coding, testing, and shipping changes in parallel, significantly reducing manual work for engineers. The initiative is reshaping developer workflows and prompting experiments with new team structures designed to manage agent fleets rather than solely human contributors.
- The internal name for Stripe's autonomous coding agents is "minions." While they are autonomous in generating code, every pull request is still reviewed by a human engineer before being merged into production systems. - This initiative builds on Stripe's long-term investment in its developer environment; the agents run in isolated, cloud-based setups called "devboxes" that were originally created to improve productivity for human engineers. These devboxes can be spun up in about 10 seconds, pre-loaded with the necessary code and services. - The core of the agent system is a customized internal fork of "goose," an open-source coding agent initially developed by Block (formerly Square). This highlights a strategy of adapting proven open-source technology rather than building entirely from scratch. - Agents are connected to Stripe's internal systems through a centralized server called "Toolshed," which provides access to over 400 different tools using the Model Context Protocol (MCP). This allows the agents to perform complex tasks by interacting with Stripe-specific libraries and services. - The system architecture is not purely LLM-driven; it's a hybrid model that combines AI-based code generation with deterministic, rule-based gates for processes like linting and running selective tests from a suite of over 3 million total tests. If an agent's fix fails tests multiple times, the task is automatically handed back to a human engineer to avoid costly, unproductive loops. - The workflow is often initiated via a Slack message, which triggers the agent to prefetch context from documentation and internal tickets before it begins coding in its sandboxed devbox environment. This "fire-and-forget" model allows engineers to delegate tasks and work on other problems in parallel.
